# Agentic Relationship Management

Ashe Magalhaes, Hearth AI | LLMs in Production 2023 | 10:01

Source: https://www.youtube.com/watch?v=U9NhEUyXKi4
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/agentic-relationship-management
Published: 2023-05-11
Tags: agents, evals, memory, workflows

## TL;DR
- Hearth AI uses agents to learn from a user's network, understand relationships, and take actions on the user's behalf.
- The relationship-warming workflow combines contact search, a relationship model, and reminders when communication has stopped.
- Production agents need benchmarks for prompts, embeddings, vector stores, memory, model interactions, and future human oversight.

## Summary
Ashe Magalhaes introduces Hearth AI's idea of agentic relationship management. The product is intended to help people handle complex personal and professional networks by enriching contact data, recommending connections, and taking actions for the user. Ashe walks through a relationship-warming workflow: a daily briefing provides context about a contact, the user asks Hearth to keep the relationship warm, and the system later sends a reminder when no communication has occurred. Behind the interface, an agent classifies the request, searches the contact knowledge base, estimates when the relationship has become cold, and runs a reminder workflow. Ashe also discusses production testing. Small prompt changes can alter outputs, chronological context can improve results, and vector-store costs require attention. As agents gain permission to act across social and work platforms, Hearth expects to need monitoring, rollback, and human review. The talk ends with Hearth's product mission and hiring plans.

## Key ideas
### Hearth AI treats network complexity as a product problem
[01:55](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=115s)
Ashe describes personal and professional networks that have become difficult to manage across communication channels. People lose track of relevant conversations, enter contact data manually, miss useful network details, and let relationships become inactive. A person attending this talk might want every attendee added to LinkedIn and Twitter, then need a way to remember who is interested in AI and when to follow up. Hearth AI's proposed category is agentic relationship management. Its products use AI agents to learn from a user's network, understand it, and act on the user's behalf so the user can spend more time on connection.

### An agent is a system that chooses observations, thoughts, and actions
[02:46](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=166s)
Ashe defines the agents behind Hearth as systems that use large language models and human feedback to decide which actions to take and in what order. The agent works toward an objective through a sequence of observations, thoughts, and actions. Hearth personalizes these systems for a person, team, or role such as sales, partnerships, business development, or capital allocation. The possible workflows include finding relationships that are going cold, identifying people to meet in a new city, preparing for a sales meeting, reporting network news, and finding people a teammate knows.

### A relationship-warming workflow turns a vague request into a scheduled reminder
[03:57](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=237s)
In Ashe's case study, Hearth sends a daily briefing about Jody before an investor introduction call. The briefing includes Jody's LinkedIn and Twitter profiles, a summary, and mutual connections. The user can tell Hearth to keep the relationship warm. If no communication occurs by a later date, Hearth sends a reminder to reach out. The user sees the Slack interaction, while the system handles the underlying work. This gives the user a simple request and a later nudge without requiring a spreadsheet or a manually maintained contact database.

### The agent routes the request through separate models and workflows
[04:48](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=288s)
For the relationship request, the agent first interprets the instruction and checks whether it matches a supported Hearth action. It searches for Jody in the knowledge base and uses the briefing or another piece of context to identify the correct person. A relationship model then estimates how much time can pass before the relationship is considered cold. Ashe says this could begin with a simple rule based on how important a person is and how much time the user has spent with them, then grow into a model of the user's preferences. A reminder workflow sends the eventual Slack message.

### Production testing must cover the whole agent system
[06:32](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=392s)
Ashe says Hearth's team is building a benchmark and test suite because many small implementation choices affect production behavior. Minor prompt changes can change output embeddings. Vector-store costs can become unexpectedly high, and the vector-store ecosystem is changing quickly as teams work with production data. Adding chronological order to meeting information produced better or more sensible outputs for Hearth. Ashe also points to the importance of self-refinement, search and action feedback, and memory that preserves user preferences over periods such as several weeks.

### Agents need monitoring as their action permissions expand
[07:36](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=456s)
Ashe describes Hearth as a suite of interacting models rather than one model. Benchmarking therefore needs to cover how those components work together. The current action layer is relatively limited because it mainly sends reminders. Future integrations with social networks and other platforms will require more monitoring, the ability to call back actions, and human review. Ashe gives automatic additions to Twitter as an example where more control would be needed. The concern is practical: once an agent can act in external systems, a mistaken action can affect a user's relationships.

### Hearth's product mission has four network functions
[08:38](https://www.youtube.com/watch?v=U9NhEUyXKi4&t=518s)
Ashe summarizes Hearth's mission as building a category of products that centralize and enrich contact information, recommend actions or connections, and act on the user's network. The purpose is to let individuals and teams focus on connecting rather than maintaining fragmented records. In the question period, Ashe says Hearth has reduced the risk of self-spam by focusing on outbound network activity and meetings with other attendees. The system had not yet reached language-model processing of emails, which limits the scope of that risk in the product described here.

## Notable quotes
- "My main mission is really to build AI that augments the human experience." (01:39)
- "Hearth AI will synthesize and manage your network complexity so that you can focus on connection." (02:28)
- "An agent's ability to self-refine and search action feedback loop is really important and how this persists in memory." (07:12)
- "The longer term actually calls for agent monitoring systems in the wild as this action layer grows." (07:47)

## Tools & references mentioned
- Hearth AI
- LinkedIn
- Twitter
- Slack
- baby AGI
- John May

## Who should watch
- You are building an agent that needs to search personal context, remember preferences, and trigger actions later.
- Your product may eventually post, message, or update records in external systems and you need to think about monitoring and human review.
- You are testing LLM workflows in production and want concrete examples of prompt sensitivity, vector-store costs, chronology, and feedback memory.

## Related talks

- [AI Agents: The Evolution of Our Digital Friends](https://mlopstalks.com/talks/ai-agents-the-evolution-of-our-digital-friends) (, 10:16)
- [How Agentic Workflows Will Change Everything](https://mlopstalks.com/talks/how-agentic-workflows-will-change-everything) (Raj Rikhy, Microsoft, 49:13)
- [The Agent Landscape - Lessons Learned Putting Agents Into Production](https://mlopstalks.com/talks/the-agent-landscape-lessons-learned-putting-agents-into-production) (Paul van der Boor & Floris Fok, Prosus Group, 1:08:41)
- [The Future of AI: Building Multi-Agent Systems](https://mlopstalks.com/talks/the-future-of-ai-building-multi-agent-systems) (, 17:30)
- [Multi-Agent Personalization with Shared Memory: From Email to Website to Proposal](https://mlopstalks.com/talks/multi-agent-personalization-with-shared-memory-from-email-to-website-to-proposal) (Hamed Taheri, Personize.ai, 16:57)